Skip to content

Model of computation#

A model of computation is a model which describes how an output of a function is computed given an input. In MLtraq, an experiment is a collection of runs that are evaluated independently by following the state monad design pattern:

"The State monad wraps computations in the context of reading and modifying a global state object. This context chains two operations together in an intuitive way. First, it determines what the state should be after the first operation. Then, it resolves the second operation with the new state."

The run is our state object, the input dictionaries run.config and run.params represent the initial state, and the step functions are the chained operations. E.g., the run after applying step functions f1, f2 and f3 in this order, is equivalent to run = f3(f2(f1(run))).


This model encourages two important design patterns:

  • Composability: Maximize reuse of existing code, design clean interfaces, separate implementation from expected outcome
  • Encapsulation: Limiting of direct access to the internal state via the state dictionaries, easying maintainability and testing


The step functions can use the state dictionaries run.vars, run.state, and run.fields to communicate and evolve the computation with varying power (and cost). In principle, one should prefer the least powerful yet sufficient semantics that get the job done.

Attribute run.vars#

It is designed to hold temporary data and is emptied once the execution of the steps is complete.

Motivation and details

With the parallel/distributed computation of runs, we want to avoid the unnecessary picking, transfer, and unpickling of objects. In some cases, objects cannot be pickled, resulting in errors and headaches. E.g., a large model of 20GB can be a good fit for run.vars if we care only about evaluation metrics. In local notebooks, we can help the garbage collector by tidying up the state of runs as soon as possible.

Attribute run.state#

It can be accessed once the run completes its execution. It is not persisted to database.

Motivation and details

It enables the inspection of the internal state of experiments for debugging and analysis after their execution. It can hold only pickable objects in case of parallel execution (parallel execution can be disabled with n_jobs=1 as per joblib.Parallel documentation). In case of unsafe pickled objects to database, its contents are persisted.

Attribute run.fields#

It works as run.state and its values are also stored in the database.

Motivation and details

It is the most robust, portable way to store the state of the experiment. With safe serialization, using native database SQL types whenever possible. Many but not all types are supported, see storage for details.

Handling of parameters#

Runs can access parameters in run.config (fixed, same for all runs) and run.params (not necessarily holding the same value for all runs). By default, these attributes are cleared after execution.

In some situations, you might want to retain access to these parameters. You can transparently request to persist them in the state. A designated run.fields attribute can be set using Experiment.execute(..., args_field="args"). If args_field is set (E.g., to "args"), their value remains accessible at any time. Under the hood, it is persisted/loaded from the dictionary field run.fields.args. The examples on Tracking demonstrate the usage of args_field to resume the computation of an experiment, including access to the parameters, after reloading it from database.

Handling of exceptions#

If an exception is raised by a step function in a run, the attribute run.exception is set and the evaluation is interrupted, reverting the run (and all the other runs of the experiment) to its initial state, triggering an exception RunException. In case of parallel/distributed execution, this mechanism ensures complete errors transparency and reporting to the driver node.

You can control the verbosity of the description of the reported exception using the option "execution.exceptions.compact_message".

Experiments as collections of aligned and evolving runs#

The management of the execution of runs at the higher abstraction level of experiments ensure that all the runs within an experiment have their inputs equally structured and the evaluation of the steps is aligned, with equal evaluation progress and consistent states.

Example: Merging runs#

Experiments are collections of runs stored in mltraq.Runs objects that extend dict, with run IDs as keys and mltraq.Run objects as values. In this example, two experiments are combined by merging their runs:

Merging runs of two experiments

from mltraq import Run, create_experiment

def step(run: Run):
    run.fields.v = run.params.v

e1 = create_experiment().add_runs(v=[1, 2]).execute(step)
e2 = create_experiment().add_runs(v=[3, 4]).execute(step)
                                 id_run  v
0  d65df69e-1175-44a5-be2f-2232765703b9  2
1  d65df69e-1175-44a5-be2f-2232765703ba  1
2  d65df69e-1175-44a5-be2f-2232765703bc  4
3  d65df69e-1175-44a5-be2f-2232765703bd  3

Handling of options#

The value of mltraq.options is transparently propagated to the context of the step functions, and can be accessed at any time in read-only mode. The options mechanism is useful to pass project configuration constants that are not necessarily mapped in run.config. For example, a step might call a load_dataset function that can be cached and the behaviour could be managed using a dedicated option. The option prefix "app." is reserved for the application, is empty by default, and can be used by the application to customize the behaviour of steps.